workqueue: fix freezeable workqueues implementation
[linux-2.6.git] / kernel / workqueue.c
1 /*
2  * linux/kernel/workqueue.c
3  *
4  * Generic mechanism for defining kernel helper threads for running
5  * arbitrary tasks in process context.
6  *
7  * Started by Ingo Molnar, Copyright (C) 2002
8  *
9  * Derived from the taskqueue/keventd code by:
10  *
11  *   David Woodhouse <dwmw2@infradead.org>
12  *   Andrew Morton <andrewm@uow.edu.au>
13  *   Kai Petzke <wpp@marie.physik.tu-berlin.de>
14  *   Theodore Ts'o <tytso@mit.edu>
15  *
16  * Made to use alloc_percpu by Christoph Lameter <clameter@sgi.com>.
17  */
18
19 #include <linux/module.h>
20 #include <linux/kernel.h>
21 #include <linux/sched.h>
22 #include <linux/init.h>
23 #include <linux/signal.h>
24 #include <linux/completion.h>
25 #include <linux/workqueue.h>
26 #include <linux/slab.h>
27 #include <linux/cpu.h>
28 #include <linux/notifier.h>
29 #include <linux/kthread.h>
30 #include <linux/hardirq.h>
31 #include <linux/mempolicy.h>
32 #include <linux/freezer.h>
33 #include <linux/kallsyms.h>
34 #include <linux/debug_locks.h>
35
36 /*
37  * The per-CPU workqueue (if single thread, we always use the first
38  * possible cpu).
39  */
40 struct cpu_workqueue_struct {
41
42         spinlock_t lock;
43
44         struct list_head worklist;
45         wait_queue_head_t more_work;
46
47         struct workqueue_struct *wq;
48         struct task_struct *thread;
49         struct work_struct *current_work;
50
51         int run_depth;          /* Detect run_workqueue() recursion depth */
52 } ____cacheline_aligned;
53
54 /*
55  * The externally visible workqueue abstraction is an array of
56  * per-CPU workqueues:
57  */
58 struct workqueue_struct {
59         struct cpu_workqueue_struct *cpu_wq;
60         const char *name;
61         struct list_head list;  /* Empty if single thread */
62         int freezeable;         /* Freeze threads during suspend */
63 };
64
65 /* All the per-cpu workqueues on the system, for hotplug cpu to add/remove
66    threads to each one as cpus come/go. */
67 static DEFINE_MUTEX(workqueue_mutex);
68 static LIST_HEAD(workqueues);
69
70 static int singlethread_cpu;
71
72 /* If it's single threaded, it isn't in the list of workqueues. */
73 static inline int is_single_threaded(struct workqueue_struct *wq)
74 {
75         return list_empty(&wq->list);
76 }
77
78 /*
79  * Set the workqueue on which a work item is to be run
80  * - Must *only* be called if the pending flag is set
81  */
82 static inline void set_wq_data(struct work_struct *work, void *wq)
83 {
84         unsigned long new;
85
86         BUG_ON(!work_pending(work));
87
88         new = (unsigned long) wq | (1UL << WORK_STRUCT_PENDING);
89         new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
90         atomic_long_set(&work->data, new);
91 }
92
93 static inline void *get_wq_data(struct work_struct *work)
94 {
95         return (void *) (atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK);
96 }
97
98 static int __run_work(struct cpu_workqueue_struct *cwq, struct work_struct *work)
99 {
100         int ret = 0;
101         unsigned long flags;
102
103         spin_lock_irqsave(&cwq->lock, flags);
104         /*
105          * We need to re-validate the work info after we've gotten
106          * the cpu_workqueue lock. We can run the work now iff:
107          *
108          *  - the wq_data still matches the cpu_workqueue_struct
109          *  - AND the work is still marked pending
110          *  - AND the work is still on a list (which will be this
111          *    workqueue_struct list)
112          *
113          * All these conditions are important, because we
114          * need to protect against the work being run right
115          * now on another CPU (all but the last one might be
116          * true if it's currently running and has not been
117          * released yet, for example).
118          */
119         if (get_wq_data(work) == cwq
120             && work_pending(work)
121             && !list_empty(&work->entry)) {
122                 work_func_t f = work->func;
123                 cwq->current_work = work;
124                 list_del_init(&work->entry);
125                 spin_unlock_irqrestore(&cwq->lock, flags);
126
127                 if (!test_bit(WORK_STRUCT_NOAUTOREL, work_data_bits(work)))
128                         work_release(work);
129                 f(work);
130
131                 spin_lock_irqsave(&cwq->lock, flags);
132                 cwq->current_work = NULL;
133                 ret = 1;
134         }
135         spin_unlock_irqrestore(&cwq->lock, flags);
136         return ret;
137 }
138
139 /**
140  * run_scheduled_work - run scheduled work synchronously
141  * @work: work to run
142  *
143  * This checks if the work was pending, and runs it
144  * synchronously if so. It returns a boolean to indicate
145  * whether it had any scheduled work to run or not.
146  *
147  * NOTE! This _only_ works for normal work_structs. You
148  * CANNOT use this for delayed work, because the wq data
149  * for delayed work will not point properly to the per-
150  * CPU workqueue struct, but will change!
151  */
152 int fastcall run_scheduled_work(struct work_struct *work)
153 {
154         for (;;) {
155                 struct cpu_workqueue_struct *cwq;
156
157                 if (!work_pending(work))
158                         return 0;
159                 if (list_empty(&work->entry))
160                         return 0;
161                 /* NOTE! This depends intimately on __queue_work! */
162                 cwq = get_wq_data(work);
163                 if (!cwq)
164                         return 0;
165                 if (__run_work(cwq, work))
166                         return 1;
167         }
168 }
169 EXPORT_SYMBOL(run_scheduled_work);
170
171 static void insert_work(struct cpu_workqueue_struct *cwq,
172                                 struct work_struct *work, int tail)
173 {
174         set_wq_data(work, cwq);
175         if (tail)
176                 list_add_tail(&work->entry, &cwq->worklist);
177         else
178                 list_add(&work->entry, &cwq->worklist);
179         wake_up(&cwq->more_work);
180 }
181
182 /* Preempt must be disabled. */
183 static void __queue_work(struct cpu_workqueue_struct *cwq,
184                          struct work_struct *work)
185 {
186         unsigned long flags;
187
188         spin_lock_irqsave(&cwq->lock, flags);
189         insert_work(cwq, work, 1);
190         spin_unlock_irqrestore(&cwq->lock, flags);
191 }
192
193 /**
194  * queue_work - queue work on a workqueue
195  * @wq: workqueue to use
196  * @work: work to queue
197  *
198  * Returns 0 if @work was already on a queue, non-zero otherwise.
199  *
200  * We queue the work to the CPU it was submitted, but there is no
201  * guarantee that it will be processed by that CPU.
202  */
203 int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)
204 {
205         int ret = 0, cpu = get_cpu();
206
207         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
208                 if (unlikely(is_single_threaded(wq)))
209                         cpu = singlethread_cpu;
210                 BUG_ON(!list_empty(&work->entry));
211                 __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
212                 ret = 1;
213         }
214         put_cpu();
215         return ret;
216 }
217 EXPORT_SYMBOL_GPL(queue_work);
218
219 void delayed_work_timer_fn(unsigned long __data)
220 {
221         struct delayed_work *dwork = (struct delayed_work *)__data;
222         struct workqueue_struct *wq = get_wq_data(&dwork->work);
223         int cpu = smp_processor_id();
224
225         if (unlikely(is_single_threaded(wq)))
226                 cpu = singlethread_cpu;
227
228         __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), &dwork->work);
229 }
230
231 /**
232  * queue_delayed_work - queue work on a workqueue after delay
233  * @wq: workqueue to use
234  * @dwork: delayable work to queue
235  * @delay: number of jiffies to wait before queueing
236  *
237  * Returns 0 if @work was already on a queue, non-zero otherwise.
238  */
239 int fastcall queue_delayed_work(struct workqueue_struct *wq,
240                         struct delayed_work *dwork, unsigned long delay)
241 {
242         int ret = 0;
243         struct timer_list *timer = &dwork->timer;
244         struct work_struct *work = &dwork->work;
245
246         timer_stats_timer_set_start_info(timer);
247         if (delay == 0)
248                 return queue_work(wq, work);
249
250         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
251                 BUG_ON(timer_pending(timer));
252                 BUG_ON(!list_empty(&work->entry));
253
254                 /* This stores wq for the moment, for the timer_fn */
255                 set_wq_data(work, wq);
256                 timer->expires = jiffies + delay;
257                 timer->data = (unsigned long)dwork;
258                 timer->function = delayed_work_timer_fn;
259                 add_timer(timer);
260                 ret = 1;
261         }
262         return ret;
263 }
264 EXPORT_SYMBOL_GPL(queue_delayed_work);
265
266 /**
267  * queue_delayed_work_on - queue work on specific CPU after delay
268  * @cpu: CPU number to execute work on
269  * @wq: workqueue to use
270  * @dwork: work to queue
271  * @delay: number of jiffies to wait before queueing
272  *
273  * Returns 0 if @work was already on a queue, non-zero otherwise.
274  */
275 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
276                         struct delayed_work *dwork, unsigned long delay)
277 {
278         int ret = 0;
279         struct timer_list *timer = &dwork->timer;
280         struct work_struct *work = &dwork->work;
281
282         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
283                 BUG_ON(timer_pending(timer));
284                 BUG_ON(!list_empty(&work->entry));
285
286                 /* This stores wq for the moment, for the timer_fn */
287                 set_wq_data(work, wq);
288                 timer->expires = jiffies + delay;
289                 timer->data = (unsigned long)dwork;
290                 timer->function = delayed_work_timer_fn;
291                 add_timer_on(timer, cpu);
292                 ret = 1;
293         }
294         return ret;
295 }
296 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
297
298 static void run_workqueue(struct cpu_workqueue_struct *cwq)
299 {
300         unsigned long flags;
301
302         /*
303          * Keep taking off work from the queue until
304          * done.
305          */
306         spin_lock_irqsave(&cwq->lock, flags);
307         cwq->run_depth++;
308         if (cwq->run_depth > 3) {
309                 /* morton gets to eat his hat */
310                 printk("%s: recursion depth exceeded: %d\n",
311                         __FUNCTION__, cwq->run_depth);
312                 dump_stack();
313         }
314         while (!list_empty(&cwq->worklist)) {
315                 struct work_struct *work = list_entry(cwq->worklist.next,
316                                                 struct work_struct, entry);
317                 work_func_t f = work->func;
318
319                 cwq->current_work = work;
320                 list_del_init(cwq->worklist.next);
321                 spin_unlock_irqrestore(&cwq->lock, flags);
322
323                 BUG_ON(get_wq_data(work) != cwq);
324                 if (!test_bit(WORK_STRUCT_NOAUTOREL, work_data_bits(work)))
325                         work_release(work);
326                 f(work);
327
328                 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
329                         printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
330                                         "%s/0x%08x/%d\n",
331                                         current->comm, preempt_count(),
332                                         current->pid);
333                         printk(KERN_ERR "    last function: ");
334                         print_symbol("%s\n", (unsigned long)f);
335                         debug_show_held_locks(current);
336                         dump_stack();
337                 }
338
339                 spin_lock_irqsave(&cwq->lock, flags);
340                 cwq->current_work = NULL;
341         }
342         cwq->run_depth--;
343         spin_unlock_irqrestore(&cwq->lock, flags);
344 }
345
346 static int worker_thread(void *__cwq)
347 {
348         struct cpu_workqueue_struct *cwq = __cwq;
349         DECLARE_WAITQUEUE(wait, current);
350         struct k_sigaction sa;
351         sigset_t blocked;
352
353         if (!cwq->wq->freezeable)
354                 current->flags |= PF_NOFREEZE;
355
356         set_user_nice(current, -5);
357
358         /* Block and flush all signals */
359         sigfillset(&blocked);
360         sigprocmask(SIG_BLOCK, &blocked, NULL);
361         flush_signals(current);
362
363         /*
364          * We inherited MPOL_INTERLEAVE from the booting kernel.
365          * Set MPOL_DEFAULT to insure node local allocations.
366          */
367         numa_default_policy();
368
369         /* SIG_IGN makes children autoreap: see do_notify_parent(). */
370         sa.sa.sa_handler = SIG_IGN;
371         sa.sa.sa_flags = 0;
372         siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
373         do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
374
375         set_current_state(TASK_INTERRUPTIBLE);
376         while (!kthread_should_stop()) {
377                 if (cwq->wq->freezeable)
378                         try_to_freeze();
379
380                 add_wait_queue(&cwq->more_work, &wait);
381                 if (list_empty(&cwq->worklist))
382                         schedule();
383                 else
384                         __set_current_state(TASK_RUNNING);
385                 remove_wait_queue(&cwq->more_work, &wait);
386
387                 if (!list_empty(&cwq->worklist))
388                         run_workqueue(cwq);
389                 set_current_state(TASK_INTERRUPTIBLE);
390         }
391         __set_current_state(TASK_RUNNING);
392         return 0;
393 }
394
395 struct wq_barrier {
396         struct work_struct      work;
397         struct completion       done;
398 };
399
400 static void wq_barrier_func(struct work_struct *work)
401 {
402         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
403         complete(&barr->done);
404 }
405
406 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
407                                         struct wq_barrier *barr, int tail)
408 {
409         INIT_WORK(&barr->work, wq_barrier_func);
410         __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
411
412         init_completion(&barr->done);
413
414         insert_work(cwq, &barr->work, tail);
415 }
416
417 static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
418 {
419         if (cwq->thread == current) {
420                 /*
421                  * Probably keventd trying to flush its own queue. So simply run
422                  * it by hand rather than deadlocking.
423                  */
424                 preempt_enable();
425                 /*
426                  * We can still touch *cwq here because we are keventd, and
427                  * hot-unplug will be waiting us to exit.
428                  */
429                 run_workqueue(cwq);
430                 preempt_disable();
431         } else {
432                 struct wq_barrier barr;
433                 int active = 0;
434
435                 spin_lock_irq(&cwq->lock);
436                 if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
437                         insert_wq_barrier(cwq, &barr, 1);
438                         active = 1;
439                 }
440                 spin_unlock_irq(&cwq->lock);
441
442                 if (active) {
443                         preempt_enable();
444                         wait_for_completion(&barr.done);
445                         preempt_disable();
446                 }
447         }
448 }
449
450 /**
451  * flush_workqueue - ensure that any scheduled work has run to completion.
452  * @wq: workqueue to flush
453  *
454  * Forces execution of the workqueue and blocks until its completion.
455  * This is typically used in driver shutdown handlers.
456  *
457  * We sleep until all works which were queued on entry have been handled,
458  * but we are not livelocked by new incoming ones.
459  *
460  * This function used to run the workqueues itself.  Now we just wait for the
461  * helper threads to do it.
462  */
463 void fastcall flush_workqueue(struct workqueue_struct *wq)
464 {
465         preempt_disable();              /* CPU hotplug */
466         if (is_single_threaded(wq)) {
467                 /* Always use first cpu's area. */
468                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, singlethread_cpu));
469         } else {
470                 int cpu;
471
472                 for_each_online_cpu(cpu)
473                         flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
474         }
475         preempt_enable();
476 }
477 EXPORT_SYMBOL_GPL(flush_workqueue);
478
479 static void wait_on_work(struct cpu_workqueue_struct *cwq,
480                                 struct work_struct *work)
481 {
482         struct wq_barrier barr;
483         int running = 0;
484
485         spin_lock_irq(&cwq->lock);
486         if (unlikely(cwq->current_work == work)) {
487                 insert_wq_barrier(cwq, &barr, 0);
488                 running = 1;
489         }
490         spin_unlock_irq(&cwq->lock);
491
492         if (unlikely(running)) {
493                 mutex_unlock(&workqueue_mutex);
494                 wait_for_completion(&barr.done);
495                 mutex_lock(&workqueue_mutex);
496         }
497 }
498
499 /**
500  * flush_work - block until a work_struct's callback has terminated
501  * @wq: the workqueue on which the work is queued
502  * @work: the work which is to be flushed
503  *
504  * flush_work() will attempt to cancel the work if it is queued.  If the work's
505  * callback appears to be running, flush_work() will block until it has
506  * completed.
507  *
508  * flush_work() is designed to be used when the caller is tearing down data
509  * structures which the callback function operates upon.  It is expected that,
510  * prior to calling flush_work(), the caller has arranged for the work to not
511  * be requeued.
512  */
513 void flush_work(struct workqueue_struct *wq, struct work_struct *work)
514 {
515         struct cpu_workqueue_struct *cwq;
516
517         mutex_lock(&workqueue_mutex);
518         cwq = get_wq_data(work);
519         /* Was it ever queued ? */
520         if (!cwq)
521                 goto out;
522
523         /*
524          * This work can't be re-queued, and the lock above protects us
525          * from take_over_work(), no need to re-check that get_wq_data()
526          * is still the same when we take cwq->lock.
527          */
528         spin_lock_irq(&cwq->lock);
529         list_del_init(&work->entry);
530         work_release(work);
531         spin_unlock_irq(&cwq->lock);
532
533         if (is_single_threaded(wq)) {
534                 /* Always use first cpu's area. */
535                 wait_on_work(per_cpu_ptr(wq->cpu_wq, singlethread_cpu), work);
536         } else {
537                 int cpu;
538
539                 for_each_online_cpu(cpu)
540                         wait_on_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
541         }
542 out:
543         mutex_unlock(&workqueue_mutex);
544 }
545 EXPORT_SYMBOL_GPL(flush_work);
546
547 static struct task_struct *create_workqueue_thread(struct workqueue_struct *wq,
548                                                         int cpu)
549 {
550         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
551         struct task_struct *p;
552
553         spin_lock_init(&cwq->lock);
554         cwq->wq = wq;
555         cwq->thread = NULL;
556         INIT_LIST_HEAD(&cwq->worklist);
557         init_waitqueue_head(&cwq->more_work);
558
559         if (is_single_threaded(wq))
560                 p = kthread_create(worker_thread, cwq, "%s", wq->name);
561         else
562                 p = kthread_create(worker_thread, cwq, "%s/%d", wq->name, cpu);
563         if (IS_ERR(p))
564                 return NULL;
565         cwq->thread = p;
566         return p;
567 }
568
569 struct workqueue_struct *__create_workqueue(const char *name,
570                                             int singlethread, int freezeable)
571 {
572         int cpu, destroy = 0;
573         struct workqueue_struct *wq;
574         struct task_struct *p;
575
576         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
577         if (!wq)
578                 return NULL;
579
580         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
581         if (!wq->cpu_wq) {
582                 kfree(wq);
583                 return NULL;
584         }
585
586         wq->name = name;
587         wq->freezeable = freezeable;
588
589         mutex_lock(&workqueue_mutex);
590         if (singlethread) {
591                 INIT_LIST_HEAD(&wq->list);
592                 p = create_workqueue_thread(wq, singlethread_cpu);
593                 if (!p)
594                         destroy = 1;
595                 else
596                         wake_up_process(p);
597         } else {
598                 list_add(&wq->list, &workqueues);
599                 for_each_online_cpu(cpu) {
600                         p = create_workqueue_thread(wq, cpu);
601                         if (p) {
602                                 kthread_bind(p, cpu);
603                                 wake_up_process(p);
604                         } else
605                                 destroy = 1;
606                 }
607         }
608         mutex_unlock(&workqueue_mutex);
609
610         /*
611          * Was there any error during startup? If yes then clean up:
612          */
613         if (destroy) {
614                 destroy_workqueue(wq);
615                 wq = NULL;
616         }
617         return wq;
618 }
619 EXPORT_SYMBOL_GPL(__create_workqueue);
620
621 static void cleanup_workqueue_thread(struct workqueue_struct *wq, int cpu)
622 {
623         struct cpu_workqueue_struct *cwq;
624         unsigned long flags;
625         struct task_struct *p;
626
627         cwq = per_cpu_ptr(wq->cpu_wq, cpu);
628         spin_lock_irqsave(&cwq->lock, flags);
629         p = cwq->thread;
630         cwq->thread = NULL;
631         spin_unlock_irqrestore(&cwq->lock, flags);
632         if (p)
633                 kthread_stop(p);
634 }
635
636 /**
637  * destroy_workqueue - safely terminate a workqueue
638  * @wq: target workqueue
639  *
640  * Safely destroy a workqueue. All work currently pending will be done first.
641  */
642 void destroy_workqueue(struct workqueue_struct *wq)
643 {
644         int cpu;
645
646         flush_workqueue(wq);
647
648         /* We don't need the distraction of CPUs appearing and vanishing. */
649         mutex_lock(&workqueue_mutex);
650         if (is_single_threaded(wq))
651                 cleanup_workqueue_thread(wq, singlethread_cpu);
652         else {
653                 for_each_online_cpu(cpu)
654                         cleanup_workqueue_thread(wq, cpu);
655                 list_del(&wq->list);
656         }
657         mutex_unlock(&workqueue_mutex);
658         free_percpu(wq->cpu_wq);
659         kfree(wq);
660 }
661 EXPORT_SYMBOL_GPL(destroy_workqueue);
662
663 static struct workqueue_struct *keventd_wq;
664
665 /**
666  * schedule_work - put work task in global workqueue
667  * @work: job to be done
668  *
669  * This puts a job in the kernel-global workqueue.
670  */
671 int fastcall schedule_work(struct work_struct *work)
672 {
673         return queue_work(keventd_wq, work);
674 }
675 EXPORT_SYMBOL(schedule_work);
676
677 /**
678  * schedule_delayed_work - put work task in global workqueue after delay
679  * @dwork: job to be done
680  * @delay: number of jiffies to wait or 0 for immediate execution
681  *
682  * After waiting for a given time this puts a job in the kernel-global
683  * workqueue.
684  */
685 int fastcall schedule_delayed_work(struct delayed_work *dwork,
686                                         unsigned long delay)
687 {
688         timer_stats_timer_set_start_info(&dwork->timer);
689         return queue_delayed_work(keventd_wq, dwork, delay);
690 }
691 EXPORT_SYMBOL(schedule_delayed_work);
692
693 /**
694  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
695  * @cpu: cpu to use
696  * @dwork: job to be done
697  * @delay: number of jiffies to wait
698  *
699  * After waiting for a given time this puts a job in the kernel-global
700  * workqueue on the specified CPU.
701  */
702 int schedule_delayed_work_on(int cpu,
703                         struct delayed_work *dwork, unsigned long delay)
704 {
705         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
706 }
707 EXPORT_SYMBOL(schedule_delayed_work_on);
708
709 /**
710  * schedule_on_each_cpu - call a function on each online CPU from keventd
711  * @func: the function to call
712  *
713  * Returns zero on success.
714  * Returns -ve errno on failure.
715  *
716  * Appears to be racy against CPU hotplug.
717  *
718  * schedule_on_each_cpu() is very slow.
719  */
720 int schedule_on_each_cpu(work_func_t func)
721 {
722         int cpu;
723         struct work_struct *works;
724
725         works = alloc_percpu(struct work_struct);
726         if (!works)
727                 return -ENOMEM;
728
729         preempt_disable();              /* CPU hotplug */
730         for_each_online_cpu(cpu) {
731                 struct work_struct *work = per_cpu_ptr(works, cpu);
732
733                 INIT_WORK(work, func);
734                 set_bit(WORK_STRUCT_PENDING, work_data_bits(work));
735                 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu), work);
736         }
737         preempt_enable();
738         flush_workqueue(keventd_wq);
739         free_percpu(works);
740         return 0;
741 }
742
743 void flush_scheduled_work(void)
744 {
745         flush_workqueue(keventd_wq);
746 }
747 EXPORT_SYMBOL(flush_scheduled_work);
748
749 void flush_work_keventd(struct work_struct *work)
750 {
751         flush_work(keventd_wq, work);
752 }
753 EXPORT_SYMBOL(flush_work_keventd);
754
755 /**
756  * cancel_rearming_delayed_workqueue - reliably kill off a delayed work whose handler rearms the delayed work.
757  * @wq:   the controlling workqueue structure
758  * @dwork: the delayed work struct
759  */
760 void cancel_rearming_delayed_workqueue(struct workqueue_struct *wq,
761                                        struct delayed_work *dwork)
762 {
763         while (!cancel_delayed_work(dwork))
764                 flush_workqueue(wq);
765 }
766 EXPORT_SYMBOL(cancel_rearming_delayed_workqueue);
767
768 /**
769  * cancel_rearming_delayed_work - reliably kill off a delayed keventd work whose handler rearms the delayed work.
770  * @dwork: the delayed work struct
771  */
772 void cancel_rearming_delayed_work(struct delayed_work *dwork)
773 {
774         cancel_rearming_delayed_workqueue(keventd_wq, dwork);
775 }
776 EXPORT_SYMBOL(cancel_rearming_delayed_work);
777
778 /**
779  * execute_in_process_context - reliably execute the routine with user context
780  * @fn:         the function to execute
781  * @ew:         guaranteed storage for the execute work structure (must
782  *              be available when the work executes)
783  *
784  * Executes the function immediately if process context is available,
785  * otherwise schedules the function for delayed execution.
786  *
787  * Returns:     0 - function was executed
788  *              1 - function was scheduled for execution
789  */
790 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
791 {
792         if (!in_interrupt()) {
793                 fn(&ew->work);
794                 return 0;
795         }
796
797         INIT_WORK(&ew->work, fn);
798         schedule_work(&ew->work);
799
800         return 1;
801 }
802 EXPORT_SYMBOL_GPL(execute_in_process_context);
803
804 int keventd_up(void)
805 {
806         return keventd_wq != NULL;
807 }
808
809 int current_is_keventd(void)
810 {
811         struct cpu_workqueue_struct *cwq;
812         int cpu = smp_processor_id();   /* preempt-safe: keventd is per-cpu */
813         int ret = 0;
814
815         BUG_ON(!keventd_wq);
816
817         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
818         if (current == cwq->thread)
819                 ret = 1;
820
821         return ret;
822
823 }
824
825 /* Take the work from this (downed) CPU. */
826 static void take_over_work(struct workqueue_struct *wq, unsigned int cpu)
827 {
828         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
829         struct list_head list;
830         struct work_struct *work;
831
832         spin_lock_irq(&cwq->lock);
833         list_replace_init(&cwq->worklist, &list);
834
835         while (!list_empty(&list)) {
836                 printk("Taking work for %s\n", wq->name);
837                 work = list_entry(list.next,struct work_struct,entry);
838                 list_del(&work->entry);
839                 __queue_work(per_cpu_ptr(wq->cpu_wq, smp_processor_id()), work);
840         }
841         spin_unlock_irq(&cwq->lock);
842 }
843
844 /* We're holding the cpucontrol mutex here */
845 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
846                                   unsigned long action,
847                                   void *hcpu)
848 {
849         unsigned int hotcpu = (unsigned long)hcpu;
850         struct workqueue_struct *wq;
851
852         switch (action) {
853         case CPU_UP_PREPARE:
854                 mutex_lock(&workqueue_mutex);
855                 /* Create a new workqueue thread for it. */
856                 list_for_each_entry(wq, &workqueues, list) {
857                         if (!create_workqueue_thread(wq, hotcpu)) {
858                                 printk("workqueue for %i failed\n", hotcpu);
859                                 return NOTIFY_BAD;
860                         }
861                 }
862                 break;
863
864         case CPU_ONLINE:
865                 /* Kick off worker threads. */
866                 list_for_each_entry(wq, &workqueues, list) {
867                         struct cpu_workqueue_struct *cwq;
868
869                         cwq = per_cpu_ptr(wq->cpu_wq, hotcpu);
870                         kthread_bind(cwq->thread, hotcpu);
871                         wake_up_process(cwq->thread);
872                 }
873                 mutex_unlock(&workqueue_mutex);
874                 break;
875
876         case CPU_UP_CANCELED:
877                 list_for_each_entry(wq, &workqueues, list) {
878                         if (!per_cpu_ptr(wq->cpu_wq, hotcpu)->thread)
879                                 continue;
880                         /* Unbind so it can run. */
881                         kthread_bind(per_cpu_ptr(wq->cpu_wq, hotcpu)->thread,
882                                      any_online_cpu(cpu_online_map));
883                         cleanup_workqueue_thread(wq, hotcpu);
884                 }
885                 mutex_unlock(&workqueue_mutex);
886                 break;
887
888         case CPU_DOWN_PREPARE:
889                 mutex_lock(&workqueue_mutex);
890                 break;
891
892         case CPU_DOWN_FAILED:
893                 mutex_unlock(&workqueue_mutex);
894                 break;
895
896         case CPU_DEAD:
897                 list_for_each_entry(wq, &workqueues, list)
898                         cleanup_workqueue_thread(wq, hotcpu);
899                 list_for_each_entry(wq, &workqueues, list)
900                         take_over_work(wq, hotcpu);
901                 mutex_unlock(&workqueue_mutex);
902                 break;
903         }
904
905         return NOTIFY_OK;
906 }
907
908 void init_workqueues(void)
909 {
910         singlethread_cpu = first_cpu(cpu_possible_map);
911         hotcpu_notifier(workqueue_cpu_callback, 0);
912         keventd_wq = create_workqueue("events");
913         BUG_ON(!keventd_wq);
914 }
915